Uncovering Insights in Coal Mill Plant Operation Modes Using GMM-LSTM-Based Unsupervised Clustering

The digital transformation of industrial processes offers new possibilities for improving plant performance and reliability. Machine learning and deep learning algorithms have been applied to monitor plant health and develop better maintenance and testing strategies. This paper focuses on the application of these techniques to a coal-fired power plant in South Africa, where coal mills play a vital role in energy generation. Detecting when the plant is running in an unsafe zone can offer insight into improving plant performance and reducing failures while moving away from routine/time-based maintenance strategies. Limited studies have focused on clustering patterns of coal mill operations combined with temporal dependencies in an unsupervised machine-learning setting. This study analyzes different temporal dependency algorithms, including Long-Short-Term-Memory, on a coal mill multivariate time-sequence dataset combined with Gaussian Mixture Models to group the operation modes of the coal mill plant. The use of GMM and LSTM/BiLSTM models on the multivariate time-sequence dataset from the coal mill mechanism shows promise in accurately clustering the different operation modes of the plant. This research demonstrates the potential of using machine learning models to solve real technical challenges in the power industry and reduce failures and move away from routine/time-based maintenance strategies.